public static Output Split(IHostEnvironment env, Input input)
{
Contracts.CheckValue(env, nameof(env));
var host = env.Register(ModuleName);
host.CheckValue(input, nameof(input));
host.Check(0 < input.Fraction && input.Fraction < 1, "The fraction must be in the interval (0,1).");
EntryPointUtils.CheckInputArgs(host, input);
var data = input.Data;
var stratCol = SplitUtils.CreateStratificationColumn(host, ref data, input.StratificationColumn);
IDataView trainData = new RangeFilter(host,
new RangeFilter.Arguments {
Column = stratCol, Min = 0, Max = input.Fraction, Complement = false
}, data);
trainData = SelectColumnsTransform.CreateDrop(host, trainData, stratCol);
IDataView testData = new RangeFilter(host,
new RangeFilter.Arguments {
Column = stratCol, Min = 0, Max = input.Fraction, Complement = true
}, data);
testData = SelectColumnsTransform.CreateDrop(host, testData, stratCol);
return(new Output()
{
TrainData = trainData, TestData = testData
});
}
public static CommonOutputs.TransformOutput RenameBinaryPredictionScoreColumns(IHostEnvironment env,
RenameBinaryPredictionScoreColumnsInput input)
{
Contracts.CheckValue(env, nameof(env));
var host = env.Register("ScoreModel");
host.CheckValue(input, nameof(input));
EntryPointUtils.CheckInputArgs(host, input);
if (input.PredictorModel.Predictor.PredictionKind == PredictionKind.BinaryClassification)
{
ColumnType labelType;
var labelNames = input.PredictorModel.GetLabelInfo(host, out labelType);
if (labelNames != null && labelNames.Length == 2)
{
var positiveClass = labelNames[1];
// Rename all the score columns.
int colMax;
var maxScoreId = input.Data.Schema.GetMaxMetadataKind(out colMax, MetadataUtils.Kinds.ScoreColumnSetId);
var copyCols = new List <(string Source, string Name)>();
for (int i = 0; i < input.Data.Schema.ColumnCount; i++)
{
if (input.Data.Schema.IsHidden(i))
{
continue;
}
if (!ShouldAddColumn(input.Data.Schema, i, null, maxScoreId))
{
continue;
}
// Do not rename the PredictedLabel column.
ReadOnlyMemory <char> tmp = default;
if (input.Data.Schema.TryGetMetadata(TextType.Instance, MetadataUtils.Kinds.ScoreValueKind, i,
ref tmp) &&
ReadOnlyMemoryUtils.EqualsStr(MetadataUtils.Const.ScoreValueKind.PredictedLabel, tmp))
{
continue;
}
var source = input.Data.Schema.GetColumnName(i);
var name = source + "." + positiveClass;
copyCols.Add((source, name));
}
var copyColumn = new CopyColumnsTransform(env, copyCols.ToArray()).Transform(input.Data);
var dropColumn = SelectColumnsTransform.CreateDrop(env, copyColumn, copyCols.Select(c => c.Source).ToArray());
return(new CommonOutputs.TransformOutput {
Model = new TransformModel(env, dropColumn, input.Data), OutputData = dropColumn
});
}
}
var newView = NopTransform.CreateIfNeeded(env, input.Data);
return(new CommonOutputs.TransformOutput {
Model = new TransformModel(env, newView, input.Data), OutputData = newView
});
}
private static IDataView UnaliasIfNeeded(IHostEnvironment env, IDataView input, KeyValuePair <string, string>[] hiddenNames)
{
if (Utils.Size(hiddenNames) == 0)
{
return(input);
}
input = CopyColumnsTransform.Create(env, new CopyColumnsTransform.Arguments()
{
Column = hiddenNames.Select(pair => new CopyColumnsTransform.Column()
{
Name = pair.Key, Source = pair.Value
}).ToArray()
}, input);
return(SelectColumnsTransform.CreateDrop(env, input, hiddenNames.Select(pair => pair.Value).ToArray()));
}
public static Output Split(IHostEnvironment env, Input input)
{
Contracts.CheckValue(env, nameof(env));
var host = env.Register(ModuleName);
host.CheckValue(input, nameof(input));
EntryPointUtils.CheckInputArgs(host, input);
var data = input.Data;
var stratCol = SplitUtils.CreateStratificationColumn(host, ref data, input.StratificationColumn);
int n = input.NumFolds;
var output = new Output
{
TrainData = new IDataView[n],
TestData = new IDataView[n]
};
// Construct per-fold datasets.
double fraction = 1.0 / n;
for (int i = 0; i < n; i++)
{
var trainData = new RangeFilter(host,
new RangeFilter.Arguments {
Column = stratCol, Min = i * fraction, Max = (i + 1) * fraction, Complement = true
}, data);
output.TrainData[i] = SelectColumnsTransform.CreateDrop(host, trainData, stratCol);
var testData = new RangeFilter(host,
new RangeFilter.Arguments {
Column = stratCol, Min = i * fraction, Max = (i + 1) * fraction, Complement = false
}, data);
output.TestData[i] = SelectColumnsTransform.CreateDrop(host, testData, stratCol);
}
return(output);
}
public static IDataTransform Create(IHostEnvironment env, Arguments args, IDataView input,
TermLoaderArguments termLoaderArgs = null)
{
Contracts.CheckValue(env, nameof(env));
var h = env.Register(LoaderSignature);
h.CheckValue(args, nameof(args));
h.CheckValue(input, nameof(input));
h.CheckUserArg(Utils.Size(args.Column) > 0, nameof(args.Column), "Columns must be specified");
// To each input column to the NgramHashExtractorArguments, a HashTransform using 31
// bits (to minimize collisions) is applied first, followed by an NgramHashTransform.
IDataView view = input;
List <TermTransform.Column> termCols = null;
if (termLoaderArgs != null)
{
termCols = new List <TermTransform.Column>();
}
var hashColumns = new List <HashTransformer.Column>();
var ngramHashColumns = new NgramHashTransform.Column[args.Column.Length];
var colCount = args.Column.Length;
// The NGramHashExtractor has a ManyToOne column type. To avoid stepping over the source
// column name when a 'name' destination column name was specified, we use temporary column names.
string[][] tmpColNames = new string[colCount][];
for (int iinfo = 0; iinfo < colCount; iinfo++)
{
var column = args.Column[iinfo];
h.CheckUserArg(!string.IsNullOrWhiteSpace(column.Name), nameof(column.Name));
h.CheckUserArg(Utils.Size(column.Source) > 0 &&
column.Source.All(src => !string.IsNullOrWhiteSpace(src)), nameof(column.Source));
int srcCount = column.Source.Length;
tmpColNames[iinfo] = new string[srcCount];
for (int isrc = 0; isrc < srcCount; isrc++)
{
var tmpName = input.Schema.GetTempColumnName(column.Source[isrc]);
tmpColNames[iinfo][isrc] = tmpName;
if (termLoaderArgs != null)
{
termCols.Add(
new TermTransform.Column
{
Name = tmpName,
Source = column.Source[isrc]
});
}
hashColumns.Add(
new HashTransformer.Column
{
Name = tmpName,
Source = termLoaderArgs == null ? column.Source[isrc] : tmpName,
HashBits = 30,
Seed = column.Seed,
Ordered = false,
InvertHash = column.InvertHash
});
}
ngramHashColumns[iinfo] =
new NgramHashTransform.Column
{
Name = column.Name,
Source = tmpColNames[iinfo],
AllLengths = column.AllLengths,
HashBits = column.HashBits,
NgramLength = column.NgramLength,
RehashUnigrams = false,
Seed = column.Seed,
SkipLength = column.SkipLength,
Ordered = column.Ordered,
InvertHash = column.InvertHash,
// REVIEW: This is an ugly internal hack to get around
// the problem that we want the *original* source names surfacing
// in the descriptions where appropriate, rather than _tmp000 and
// what have you. The alternative is we do something elaborate
// with metadata or something but I'm not sure that's better.
FriendlyNames = column.FriendlyNames
};
}
if (termLoaderArgs != null)
{
h.Assert(Utils.Size(termCols) == hashColumns.Count);
var termArgs =
new TermTransform.Arguments()
{
MaxNumTerms = int.MaxValue,
Terms = termLoaderArgs.Terms,
Term = termLoaderArgs.Term,
DataFile = termLoaderArgs.DataFile,
Loader = termLoaderArgs.Loader,
TermsColumn = termLoaderArgs.TermsColumn,
Sort = termLoaderArgs.Sort,
Column = termCols.ToArray()
};
view = TermTransform.Create(h, termArgs, view);
if (termLoaderArgs.DropUnknowns)
{
var naDropArgs = new NADropTransform.Arguments {
Column = new NADropTransform.Column[termCols.Count]
};
for (int iinfo = 0; iinfo < termCols.Count; iinfo++)
{
naDropArgs.Column[iinfo] =
new NADropTransform.Column {
Name = termCols[iinfo].Name, Source = termCols[iinfo].Name
};
}
view = new NADropTransform(h, naDropArgs, view);
}
}
// Args for the Hash function with multiple columns
var hashArgs =
new HashTransformer.Arguments
{
HashBits = 31,
Seed = args.Seed,
Ordered = false,
Column = hashColumns.ToArray(),
InvertHash = args.InvertHash
};
view = HashTransformer.Create(h, hashArgs, view);
// creating the NgramHash function
var ngramHashArgs =
new NgramHashTransform.Arguments
{
AllLengths = args.AllLengths,
HashBits = args.HashBits,
NgramLength = args.NgramLength,
SkipLength = args.SkipLength,
RehashUnigrams = false,
Ordered = args.Ordered,
Seed = args.Seed,
Column = ngramHashColumns,
InvertHash = args.InvertHash
};
view = new NgramHashTransform(h, ngramHashArgs, view);
return(SelectColumnsTransform.CreateDrop(h, view, tmpColNames.SelectMany(cols => cols).ToArray()));
}
public static IDataTransform Create(IHostEnvironment env, Arguments args, IDataView input)
{
Contracts.CheckValue(env, nameof(env));
var h = env.Register(RegistrationName);
h.CheckValue(args, nameof(args));
h.CheckValue(input, nameof(input));
h.CheckUserArg(Utils.Size(args.Column) > 0, nameof(args.Column), "Columns must be specified");
// To each input column to the WordHashBagTransform, a tokenize transform is applied,
// followed by applying WordHashVectorizeTransform.
// Since WordHashBagTransform is a many-to-one column transform, for each
// WordHashBagTransform.Column we may need to define multiple tokenize transform columns.
// NgramHashExtractorTransform may need to define an identical number of HashTransform.Columns.
// The intermediate columns are dropped at the end of using a DropColumnsTransform.
IDataView view = input;
var uniqueSourceNames = NgramExtractionUtils.GenerateUniqueSourceNames(h, args.Column, view.Schema);
Contracts.Assert(uniqueSourceNames.Length == args.Column.Length);
var tokenizeColumns = new List <WordTokenizeTransform.ColumnInfo>();
var extractorCols = new NgramHashExtractorTransform.Column[args.Column.Length];
var colCount = args.Column.Length;
List <string> tmpColNames = new List <string>();
for (int iinfo = 0; iinfo < colCount; iinfo++)
{
var column = args.Column[iinfo];
int srcCount = column.Source.Length;
var curTmpNames = new string[srcCount];
Contracts.Assert(uniqueSourceNames[iinfo].Length == args.Column[iinfo].Source.Length);
for (int isrc = 0; isrc < srcCount; isrc++)
{
tokenizeColumns.Add(new WordTokenizeTransform.ColumnInfo(args.Column[iinfo].Source[isrc], curTmpNames[isrc] = uniqueSourceNames[iinfo][isrc]));
}
tmpColNames.AddRange(curTmpNames);
extractorCols[iinfo] =
new NgramHashExtractorTransform.Column
{
Name = column.Name,
Source = curTmpNames,
HashBits = column.HashBits,
NgramLength = column.NgramLength,
Seed = column.Seed,
SkipLength = column.SkipLength,
Ordered = column.Ordered,
InvertHash = column.InvertHash,
FriendlyNames = args.Column[iinfo].Source,
AllLengths = column.AllLengths
};
}
view = new WordTokenizingEstimator(env, tokenizeColumns.ToArray()).Fit(view).Transform(view);
var featurizeArgs =
new NgramHashExtractorTransform.Arguments
{
AllLengths = args.AllLengths,
HashBits = args.HashBits,
NgramLength = args.NgramLength,
SkipLength = args.SkipLength,
Ordered = args.Ordered,
Seed = args.Seed,
Column = extractorCols.ToArray(),
InvertHash = args.InvertHash
};
view = NgramHashExtractorTransform.Create(h, featurizeArgs, view);
// Since we added columns with new names, we need to explicitly drop them before we return the IDataTransform.
return(SelectColumnsTransform.CreateDrop(h, view, tmpColNames.ToArray()));
}
public void NormalizerWorkout()
{
string dataPath = GetDataPath("iris.txt");
var loader = new TextLoader(Env, new TextLoader.Arguments
{
Column = new[] {
new TextLoader.Column("float1", DataKind.R4, 1),
new TextLoader.Column("float4", DataKind.R4, new[] { new TextLoader.Range(1, 4) }),
new TextLoader.Column("double1", DataKind.R8, 1),
new TextLoader.Column("double4", DataKind.R8, new[] { new TextLoader.Range(1, 4) }),
new TextLoader.Column("int1", DataKind.I4, 0),
new TextLoader.Column("float0", DataKind.R4, new[] { new TextLoader.Range {
Min = 1, VariableEnd = true
} }),
},
HasHeader = true
}, new MultiFileSource(dataPath));
var est = new NormalizingEstimator(Env,
new NormalizingEstimator.MinMaxColumn("float1"),
new NormalizingEstimator.MinMaxColumn("float4"),
new NormalizingEstimator.MinMaxColumn("double1"),
new NormalizingEstimator.MinMaxColumn("double4"),
new NormalizingEstimator.BinningColumn("float1", "float1bin"),
new NormalizingEstimator.BinningColumn("float4", "float4bin"),
new NormalizingEstimator.BinningColumn("double1", "double1bin"),
new NormalizingEstimator.BinningColumn("double4", "double4bin"),
new NormalizingEstimator.MeanVarColumn("float1", "float1mv"),
new NormalizingEstimator.MeanVarColumn("float4", "float4mv"),
new NormalizingEstimator.MeanVarColumn("double1", "double1mv"),
new NormalizingEstimator.MeanVarColumn("double4", "double4mv"),
new NormalizingEstimator.LogMeanVarColumn("float1", "float1lmv"),
new NormalizingEstimator.LogMeanVarColumn("float4", "float4lmv"),
new NormalizingEstimator.LogMeanVarColumn("double1", "double1lmv"),
new NormalizingEstimator.LogMeanVarColumn("double4", "double4lmv"));
var data = loader.Read(dataPath);
var badData1 = new CopyColumnsTransform(Env, ("int1", "float1")).Transform(data);
var badData2 = new CopyColumnsTransform(Env, ("float0", "float4")).Transform(data);
TestEstimatorCore(est, data, null, badData1);
TestEstimatorCore(est, data, null, badData2);
var outputPath = GetOutputPath("NormalizerEstimator", "normalized.tsv");
using (var ch = Env.Start("save"))
{
var saver = new TextSaver(Env, new TextSaver.Arguments {
Silent = true
});
using (var fs = File.Create(outputPath))
{
var dataView = SelectColumnsTransform.CreateDrop(Env, est.Fit(data).Transform(data), "float0");
DataSaverUtils.SaveDataView(ch, saver, dataView, fs, keepHidden: true);
}
}
CheckEquality("NormalizerEstimator", "normalized.tsv");
Done();
}
public static IDataTransform Create(IHostEnvironment env, Arguments args, IDataView input)
{
Contracts.CheckValue(env, nameof(env));
var h = env.Register(LoaderSignature);
h.CheckValue(args, nameof(args));
h.CheckValue(input, nameof(input));
h.CheckNonWhiteSpace(args.Source, nameof(args.Source));
if (string.IsNullOrWhiteSpace(args.Name))
{
args.Name = args.Source;
}
var file = Utils.FindExistentFileOrNull("pretrained.model", "Sentiment", assemblyForBasePath: typeof(SentimentAnalyzingTransform));
if (file == null)
{
throw h.Except("resourcePath", "Missing resource for SentimentAnalyzingTransform.");
}
// The logic below ensures that any columns in our input IDataView that conflict
// with column names known to be used in the pretrained model transform pipeline we're
// loading are aliased to temporary column names before we apply the pipeline and then
// renamed back to their original names after. We do this to ensure the pretrained model
// doesn't shadow or replace columns we aren't expecting it to.
// 1. Alias any column in the input IDataView that is known to appear to the pretrained
// model into a temporary column so that we can restore them after the pretrained model
// is added to the pipeline.
KeyValuePair <string, string>[] aliased;
input = AliasIfNeeded(env, input, _modelIntermediateColumnNames, out aliased);
// 2. Copy source column to a column with the name expected by the pretrained model featurization
// transform pipeline.
var copyTransformer = new CopyColumnsTransform(env, (args.Source, ModelInputColumnName));
input = copyTransformer.Transform(input);
// 3. Apply the pretrained model and its featurization transform pipeline.
input = LoadTransforms(env, input, file);
// 4. Copy the output column from the pretrained model to a temporary column.
var scoreTempName = input.Schema.GetTempColumnName("sa_out");
copyTransformer = new CopyColumnsTransform(env, (ModelScoreColumnName, scoreTempName));
input = copyTransformer.Transform(input);
// 5. Drop all the columns created by the pretrained model, including the expected input column
// and the output column, which we have copied to a temporary column in (4).
input = SelectColumnsTransform.CreateDrop(env, input, _modelIntermediateColumnNames);
// 6. Unalias all the original columns that were originally present in the IDataView, but may have
// been shadowed by column names in the pretrained model. This method will also drop all the temporary
// columns that were created for them in (1).
input = UnaliasIfNeeded(env, input, aliased);
// 7. Copy the temporary column with the score we created in (4) to a column with the user-specified destination name.
copyTransformer = new CopyColumnsTransform(env, (scoreTempName, args.Name));
input = copyTransformer.Transform(input);
// 8. Drop the temporary column with the score created in (4).
return(SelectColumnsTransform.CreateDrop(env, input, scoreTempName));
}